Ore image segmentation by learning image and shape features Dipti Prasad Mukherjee * , Yury Potapovich, Ilya Levner, Hong Zhang Centre for Intelligent Mining Systems, Department of Computing Science, University of Alberta, Canada article info Article history: Received 12 November 2007 Received in revised form 24 November 2008 Available online 9 January 2009 Communicated by H. Sako Keywords: Ore fragment segmentation Ore size analysis Machine learning abstract We present an image segmentation system specifically targeted for oil sand ore size estimation. The sys- tem learns spectral and shape characteristics of training images of oil sand ore samples for image quality enhancement followed by segmentation of ore image shapes. The proposed segmentation has achieved superior accuracy over the current state of the art systems. Ó 2009 Elsevier B.V. All rights reserved. 1. Introduction Ore size is an important indicator for process performance in mining industry, especially for oil sand mining for extraction of crude oil. Use of image segmentation for ore size analysis is an established tool. A limited success has already been achieved with a few commercially available systems like (Girdner et al., 1996; Maerz et al., 1996). However, adaptation of such systems for oil sand is an extremely difficult task mainly because of the inhospita- ble environment common to oil sand mines. Given the uncon- trolled lighting and variability of ore surface characteristics, it is indeed a challenging task to segment the objects of interest from the images of oil sand fragments. In this paper we present an image processing system that can segment oil sand images to a significant accuracy. A typical oil sand image is shown in Fig. 1a. It is observed that for a typical oil sand image, pixel groups belonging to a single ob- ject of interest can have completely different spectral and/or tex- tural characteristics. This is due to uneven oil sand fragment surface, high speed motion of the fragments on a dirt-filled con- veyer belt and above all due to lighting conditions influenced by open-air weather conditions. Therefore, a need exists to train the image processing system such that pixel clusters having different spectral characteristics can still be identified as being part of one oil sand fragment. Otherwise, any unsupervised approach of image segmentation would definitely cluster these pixel groups into sep- arate regions. Like any typical image segmentation system, the proposed sys- tem has two definite stages, viz. preprocessing and segmentation. The preprocessing is dedicated to enhance the quality of oil sand image. However, this is implemented in a machine learning frame- work where image enhancement is motivated by the typical spec- tral characteristics of oil sand fragment images known a priori. These image characteristics are learned using logistic regression technique (Hastie et al., 2001), which uses a large number of spec- tral and textural features in order to achieve a fair amount of generality. The segmentation technique generates binary blob shapes through image thresholding at all image intensity level. The proce- dure ensures that all possible binary blobs are being investigated for shape properties like area, convexity, eccentricity or solidity common to oil sand fragment shapes (Mukherjee et al., 2004). This process of segmentation recovers exact shape or fragment bound- ary as opposed to distorted shape, which is obtained when, for example, a structuring element of a particular shape is imposed on the blob representing a fragment in morphology based image segmentation technique (Dornaika and Zhang, 2000). This we claim is an important advantage of the proposed technique where determination of exact ore size is the goal. Similar to learning image spectral properties, possible oil sand fragment shape measures can also be learned in order to automate the ore detection process. This is more so as our ultimate goal is to use the proposed system in an industrial setting where minimum number of parameters should be left for the user to tune. We have used regression-based classifier to learn the shape characteristics, like the possible range of area, solidity and eccentricity of a poten- tial oil sand fragment (Duda et al., 2001). The proposed segmentation algorithm, once tested and vali- dated, can be installed in the field to provide real-time ore size measurement for the industrial sponsor of this research as well as for other companies in the oil sands industry, for the purpose 0167-8655/$ - see front matter Ó 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.patrec.2008.12.015 * Corresponding author. Tel.: +1 91 33 577 8049; fax: +1 91 332 577 3035. E-mail address: dipti@isical.ac.in (D.P. Mukherjee). Pattern Recognition Letters 30 (2009) 615–622 Contents lists available at ScienceDirect Pattern Recognition Letters journal homepage: www.elsevier.com/locate/patrec